[2/2] net/mlx5: optimize mprq memcpy for AMD EPYC2 platforms

Message ID 20210823084411.29592-2-aman.kumar@vvdntech.in (mailing list archive)
State Superseded, archived
Delegated to: Thomas Monjalon
Headers
Series [1/2] lib/eal: add amd epyc2 memcpy routine to eal |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/github-robot: build success github build: passed
ci/iol-broadcom-Performance success Performance Testing PASS
ci/iol-broadcom-Functional success Functional Testing PASS
ci/iol-intel-Functional success Functional Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/iol-aarch64-compile-testing success Testing PASS
ci/iol-mellanox-Performance success Performance Testing PASS
ci/Intel-compilation success Compilation OK
ci/intel-Testing success Testing PASS
ci/iol-x86_64-unit-testing success Testing PASS
ci/iol-x86_64-compile-testing success Testing PASS
ci/iol-aarch64-unit-testing fail Testing issues

Commit Message

Aman Kumar Aug. 23, 2021, 8:44 a.m. UTC
add non temporal load and temporal store for mprq memcpy.
define mlx5_ntload_tstore in meson build configuration to
enable this optimization. This utilizes AMD EPYC2 optimized
rte_memcpy* routines.

Signed-off-by: Aman Kumar <aman.kumar@vvdntech.in>
---
 drivers/net/mlx5/meson.build |  1 +
 drivers/net/mlx5/mlx5.c      | 12 ++++++++++++
 drivers/net/mlx5/mlx5.h      |  3 +++
 drivers/net/mlx5/mlx5_rx.h   | 24 ++++++++++++++++++++++++
 drivers/net/mlx5/mlx5_rxq.c  |  4 ++++
 meson_options.txt            |  2 ++
 6 files changed, 46 insertions(+)
  

Comments

Thomas Monjalon Oct. 13, 2021, 4:53 p.m. UTC | #1
23/08/2021 10:44, Aman Kumar:
> add non temporal load and temporal store for mprq memcpy.
> define mlx5_ntload_tstore in meson build configuration to
> enable this optimization. This utilizes AMD EPYC2 optimized
> rte_memcpy* routines.
[...]
> +option('mlx5_ntload_tstore', type: 'boolean', value: false, description:
> +       'to enable optimized MPRQ in RX datapath')

Please don't make it a compilation option.
Why isn't it always enabled?

There was a comment on the first patch.
Do you plan to make a new version?
  
Aman Kumar Oct. 19, 2021, 10:52 a.m. UTC | #2
Dear Thomas,

Please check V2 patchset
<https://patchwork.dpdk.org/project/dpdk/patch/20211019104724.19416-1-aman.kumar@vvdntech.in/>.
Upon suggestions from Jerin, we've moved this build option using
config/x86/x86_amd_epyc_linux_gcc cross-file.
This option is not a compilation option and enabled by default in V2, once
it is cross built. Please let us know your comments.


*With Best Regards*
Aman Kumar
VVDN Technologies Pvt. Ltd.
*web:* www.vvdntech.com




On Wed, Oct 13, 2021 at 10:23 PM Thomas Monjalon <thomas@monjalon.net>
wrote:

> 23/08/2021 10:44, Aman Kumar:
> > add non temporal load and temporal store for mprq memcpy.
> > define mlx5_ntload_tstore in meson build configuration to
> > enable this optimization. This utilizes AMD EPYC2 optimized
> > rte_memcpy* routines.
> [...]
> > +option('mlx5_ntload_tstore', type: 'boolean', value: false, description:
> > +       'to enable optimized MPRQ in RX datapath')
>
> Please don't make it a compilation option.
> Why isn't it always enabled?
>
> There was a comment on the first patch.
> Do you plan to make a new version?
>
>
>
  

Patch

diff --git a/drivers/net/mlx5/meson.build b/drivers/net/mlx5/meson.build
index dac7f1fabf..0d2888742c 100644
--- a/drivers/net/mlx5/meson.build
+++ b/drivers/net/mlx5/meson.build
@@ -61,6 +61,7 @@  foreach option:cflags_options
         cflags += option
     endif
 endforeach
+dpdk_conf.set('RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY', get_option('mlx5_ntload_tstore'))
 if get_option('buildtype').contains('debug')
     cflags += [ '-pedantic', '-DPEDANTIC' ]
 else
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index f84e061fe7..cf57867c25 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -161,6 +161,11 @@ 
 /* Configure timeout of LRO session (in microseconds). */
 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
 
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+/* mprq_tstore_memcpy */
+#define MLX5_MPRQ_TSTORE_MEMCPY "mprq_tstore_memcpy"
+#endif
+
 /*
  * Device parameter to configure the total data buffer size for a single
  * hairpin queue (logarithm value).
@@ -1991,6 +1996,10 @@  mlx5_args_check(const char *key, const char *val, void *opaque)
 		config->decap_en = !!tmp;
 	} else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
 		config->allow_duplicate_pattern = !!tmp;
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	} else if (strcmp(MLX5_MPRQ_TSTORE_MEMCPY, key) == 0) {
+		config->mprq_tstore_memcpy = tmp;
+#endif
 	} else {
 		DRV_LOG(WARNING, "%s: unknown parameter", key);
 		rte_errno = EINVAL;
@@ -2051,6 +2060,9 @@  mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 		MLX5_SYS_MEM_EN,
 		MLX5_DECAP_EN,
 		MLX5_ALLOW_DUPLICATE_PATTERN,
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+		MLX5_MPRQ_TSTORE_MEMCPY,
+#endif
 		NULL,
 	};
 	struct rte_kvargs *kvlist;
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index e02714e231..7d5617f5ca 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -298,6 +298,9 @@  struct mlx5_dev_config {
 	int tx_skew; /* Tx scheduling skew between WQE and data on wire. */
 	struct mlx5_hca_attr hca_attr; /* HCA attributes. */
 	struct mlx5_lro_config lro; /* LRO configuration. */
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	unsigned int mprq_tstore_memcpy:1;
+#endif
 };
 
 
diff --git a/drivers/net/mlx5/mlx5_rx.h b/drivers/net/mlx5/mlx5_rx.h
index 3f2b99fb65..19318bdd1b 100644
--- a/drivers/net/mlx5/mlx5_rx.h
+++ b/drivers/net/mlx5/mlx5_rx.h
@@ -148,6 +148,9 @@  struct mlx5_rxq_data {
 	uint32_t rxseg_n; /* Number of split segment descriptions. */
 	struct mlx5_eth_rxseg rxseg[MLX5_MAX_RXQ_NSEG];
 	/* Buffer split segment descriptions - sizes, offsets, pools. */
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	unsigned int mprq_tstore_memcpy:1;
+#endif
 } __rte_cache_aligned;
 
 enum mlx5_rxq_type {
@@ -422,6 +425,15 @@  mprq_buf_to_pkt(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt, uint32_t len,
 	const uint32_t offset = strd_idx * strd_sz + strd_shift;
 	void *addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset);
 
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	if (unlikely(!rxq->mprq_tstore_memcpy) &&
+	    len <= rxq->mprq_max_memcpy_len) {
+		rte_prefetch1(addr);
+		if (len > RTE_CACHE_LINE_SIZE)
+			rte_prefetch2((void *)((uintptr_t)addr +
+					       RTE_CACHE_LINE_SIZE));
+	}
+#endif
 	/*
 	 * Memcpy packets to the target mbuf if:
 	 * - The size of packet is smaller than mprq_max_memcpy_len.
@@ -433,8 +445,20 @@  mprq_buf_to_pkt(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt, uint32_t len,
 	    (hdrm_overlap > 0 && !rxq->strd_scatter_en)) {
 		if (likely(len <=
 			   (uint32_t)(pkt->buf_len - RTE_PKTMBUF_HEADROOM))) {
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+			uintptr_t data_addr;
+
+			data_addr = (uintptr_t)rte_pktmbuf_mtod(pkt, void *);
+			if (!((data_addr | (uintptr_t)addr) & ALIGNMENT_MASK) &&
+				rxq->mprq_tstore_memcpy)
+				rte_memcpy_aligned_tstore16((void *)data_addr,
+					   addr, len);
+			else
+				rte_memcpy((void *)data_addr, addr, len);
+#else
 			rte_memcpy(rte_pktmbuf_mtod(pkt, void *),
 				   addr, len);
+#endif
 			DATA_LEN(pkt) = len;
 		} else if (rxq->strd_scatter_en) {
 			struct rte_mbuf *prev = pkt;
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index abd8ce7989..a1b0fa6455 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -1449,6 +1449,10 @@  mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 	tmpl->socket = socket;
 	if (dev->data->dev_conf.intr_conf.rxq)
 		tmpl->irq = 1;
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	tmpl->rxq.mprq_tstore_memcpy = config->mprq_tstore_memcpy;
+#endif
+
 	/*
 	 * This Rx queue can be configured as a Multi-Packet RQ if all of the
 	 * following conditions are met:
diff --git a/meson_options.txt b/meson_options.txt
index e232c9c340..cc7b629d17 100644
--- a/meson_options.txt
+++ b/meson_options.txt
@@ -38,6 +38,8 @@  option('max_lcores', type: 'integer', value: 128, description:
        'maximum number of cores/threads supported by EAL')
 option('max_numa_nodes', type: 'integer', value: 32, description:
        'maximum number of NUMA nodes supported by EAL')
+option('mlx5_ntload_tstore', type: 'boolean', value: false, description:
+       'to enable optimized MPRQ in RX datapath')
 option('platform', type: 'string', value: 'native', description:
        'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.')
 option('enable_trace_fp', type: 'boolean', value: false, description: